GuideWeb: leaderboard
GuideWeb (Gan et al., 2026) asks a model to read the raw HTML of a real website's main page, pick the interactive elements that merit an in-app guide and write the guide text, checked against human-verified guides for held-out pages. F1 in percent between the chosen and the reference guide elements, with the model prompted zero-shot; higher is better.
Metric: Guide-target selection F1 (%; predicted guide elements against human-verified gold DOM elements, zero-shot prompt over the raw page HTML). Source: arxiv.org. Saturation forecast: Around February 2028. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.5 | 25.16 | #244 |
| 2 | GPT-5 | 23.9 | #163 |
| 3 | Gemini 2.5 Pro | 22.85 | #262 |
| 4 | Qwen 3 8B | 10.48 | #1838 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=guideweb · How It Works · Data refreshed daily, snapshot 2026-11-12.